# Rethinking Layer-wise Feature Amounts in Convolutional Neural Network Architectures

## TL;DR

This paper challenges the common assumption that CNNs should have increasing feature counts in higher layers, showing that larger early layers can improve accuracy based on experiments with VGG-like models.

## Contribution

It introduces a skew normal distribution framework to analyze feature distributions across CNN layers and suggests rethinking layer-wise feature allocation strategies.

## Key findings

- Larger early layers can lead to better accuracy in CNNs.
- The common assumption of increasing features in higher layers is not always optimal.
- Experimental evidence from VGG-type models on multiple datasets supports this view.

## Abstract

We characterize convolutional neural networks with respect to the relative amount of features per layer. Using a skew normal distribution as a parametrized framework, we investigate the common assumption of monotonously increasing feature-counts with higher layers of architecture designs. Our evaluation on models with VGG-type layers on the MNIST, Fashion-MNIST and CIFAR-10 image classification benchmarks provides evidence that motivates rethinking of our common assumption: architectures that favor larger early layers seem to yield better accuracy.